Date Added: Mar 2011
A new compilation framework enables the execution of numerical-intensive applications in an execution environment that is formed by multi-core Central Processing Units (CPUs) and Graphics Processing Units (GPUs). A critical innovation is the use of a variation of Linear Memory Access Descriptors (LMADs) to analyze loop nests and determine automatically which memory locations must be transferred between the CPU address space and the GPU address space. In this programming model, the application is written in a combination of Python and NumPy, a rich numerical extension for Python. Inobstrusive light annotation is introduced to identify the type of function parameters and return values, and to indicate which loop nests should be parallelized and executed in the GPU.